Multilevel modeling of complex survey data

Survey data are often analyzed using multilevel or hierarchical models. For example, in education surveys, schools may be sampled at the first stage and students at the second stage and multilevel models used to model within-school and between-school variability. An important aspect of most surveys that is often ignored in multilevel modeling is that units at each stage are sampled with unequal probabilities. Standard maximum likelihood estimation can be modified to take the sampling probabilities into account, yielding pseudomaximum likelihood estimation, which is typically combined with robust standard errors based on the sandwich estimator. This approach is implemented in gllamm. I will introduce the ideas, discuss issues that arise such as the scaling of the weights, and illustrate the approach by applying it to data from the Program for International Student Assessment (PISA).